The Chip Sovereignty Myth: Why the $280B Fab Race Is Building Yesterday’s Advantage

Governments are spending hundreds of billions building chip factories — and the spending may be answering the wrong question.

That is the reframe AMD chief technology officer Mark Papermaster offered in an EE Times interview published this week: AI advantage no longer tracks to any single manufacturing node or fab. A year ago, he said, AI was associated with a GPU. Over the last year, it has become clear that you need extensive computation across a much broader system — CPUs, accelerators, memory bandwidth, packaging, and software working together. No single chip, no single fab, no single country controls all of it.

The timing matters. Policymakers in Washington, Brussels, Tokyo, and New Delhi have committed what analysts estimate at roughly $280 billion in semiconductor subsidy programs globally — drawn from reported national commitments including the CHIPS Act's roughly $52 billion for American fabrication, Europe's roughly $47 billion program, and comparable initiatives across Asia. The assumption behind much of it: build the fabs, capture the AI future.

The Papermaster interview is not long — a few hundred words — but it lands at a moment when the policy consensus is hardening around a version of the problem that the industry itself has moved past.

The chokepoint nobody is funding

The real bottleneck in AI compute today is not leading-edge wafer capacity. It is advanced chip packaging — the step where completed dies are assembled into working systems that can actually do useful work. TSMC accounts for about 90% of advanced packaging for AI chips. Nvidia has secured roughly 60% of that capacity, straining access for Google, AMD, and Microsoft.

Advanced packaging — the techniques that integrate multiple chips into a single module, connect high-bandwidth memory to processors, and manage thermal loads in dense compute systems — is what makes AI chips actually performant at scale. Without it, a leading-edge wafer is silicon waiting for assembly.

TSMC pledged an additional $100 billion last year to build advanced packaging capability in the United States, bringing its total Arizona commitment to $165 billion. TSMC has not announced when it will begin construction on those packaging facilities. None exist yet. Arizona fabs are shipping finished wafers back to Taiwan for packaging because there is nowhere else to do it at scale.

The policy debate remains fixed on wafer fabrication. Packaging — the step that actually determines whether a system can sustain modern AI workloads — is not on the subsidy map.

The oil parallel

The pattern has precedent. In the 1970s, OPEC countries nationalized oil production and seized control of wells and reservoirs. The strategic assumption: ownership of the physical resource translated to economic and geopolitical power. What it delivered was leverage over crude oil prices while the global economy ran on refined products, pipeline networks, and petrochemical infrastructure that remained under private and Western control. The producers owned the ground. The consumers controlled everything that happened after the oil came out of it.

Semiconductors are tracking the same arc. Wafer fabrication has become the focus of national industrial policy. The layer that actually determines whether a chip can perform at scale — advanced packaging — is concentrated almost entirely in Taiwan and South Korea, remains largely unaddressed by subsidy programs, and represents the true chokepoint in the AI supply chain.

Chris Miller, author of Chip War and a professor at Tufts University, frames the issue around what he calls the hidden chokepoints below the chip level. If you dig into supply chains, what you find is that many of the most vulnerable points are actually in the subcomponents and the materials, he told EE Times. These are not where the most economic value is captured, but they determine whether entire segments of the industry function at all.

The CHIPS Act as written funds wafer fabs. It does not meaningfully address packaging.

Europe misreads itself

Miller also offers a useful corrective to the prevailing European narrative. Policymakers in Brussels often look at the absence of leading-edge fabs and conclude that Europe has a weak semiconductor ecosystem. In fact, the opposite is true. Europe is a strong player in tools, materials, and advanced research. The weakness framing reflects a narrow definition of what semiconductor strength looks like.

US firms have long led in system architecture, chip design, and software rather than manufacturing. That positioning — which AMD and Nvidia exploit through chiplet architectures and heterogeneous integration — plays to US strengths in a system-centric world. The countries building leading-edge fabs may be building capacity that is necessary but no longer sufficient.

What the money is actually buying

The global semiconductor subsidy wave will produce real fabs with real leading-edge capacity. TSMC Arizona will fab chips. Samsung and Intel will add domestic capability. These are not trivial outcomes.

But the investment is calibrated to the 2020 version of the problem, when the concern was chip supply concentration and the answer was domestic fabrication. The 2026 version is different: AI compute advantage now depends on system integration — how chips, memory, packaging, and software combine. The chokepoint that actually constrains AI scaling is not the wafer that gets etched. It is the packaging step that determines whether a system can sustain the thermal and bandwidth demands of modern AI workloads.

No country can replicate the entire semiconductor ecosystem within its own borders, Papermaster said. Resilience comes from diversity and openness across the supply chain. That is not a comfortable message for policymakers who have bet hundreds of billions on a simpler answer.

Governments are building a fence around yesterday's bottleneck. The real one is somewhere else, and nobody with a national treasury is building a fence around that.